Abstract
OBJECTIVE: This study aims to automatically classify lung conditions into normal, non-small cell lung cancer (NSCLC), and small cell lung cancer (SCLC) using [(18)F] FDG PET/CT images and deep learning. METHODS: PET/CT scans from 146 patients (1974 scans) were retrospectively analyzed using two strategies: (1) transfer learning with pre-trained CNNs, and (2) a custom CNN (Res-SE Net) incorporating residual and squeeze-and-excitation (SE) modules. A patient-based data splitting approach was used to avoid data leakage. Models were trained and validated at the scan level and evaluated at the patient level using majority voting. Grad-CAM was employed to generate lesion-localization heatmaps. RESULTS: Among the seven evaluated CNN models, the proposed Res-SE Net demonstrated superior performance, achieving an accuracy of 91.67% and a sensitivity of 92.00% in detecting NSCLC, and an accuracy of 90.14% with a sensitivity of 90.00% for distinguishing SCLC cases. When tested on an external dataset, the model attained an accuracy of 98.00% in binary classification (Normal vs. Cancer). In the three-class classification task, the model achieved an accuracy of 73.02% for NSCLC and 66.26% for SCLC. CONCLUSION: These findings demonstrate the potential of Res-SE Net architecture for accurate multi-class lung cancer classification using [18F] FDG PET/CT images.